Reproduce “Self Normalizing Flows” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
T. Anderson Keller, Jorn W.T. Peters, Priyank Jaini, Emiel Hoogeboom, Patrick Forré, Max Welling
SNF reduces flow-training costs by learning an approximate inverse for each layer and using it to estimate the costly likelihood-gradient term. A reconstruction loss aligns forward and inverse maps, while reported likelihoods remain exact. On two-layer fully connected MNIST, SNF achieved 947.1 test nats and 18.7 ms per batch, versus 947.6 and 44.9 ms with exact gradients; this evidence is limited to tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.